C++AMP on LinuxMiller Lee

About Me

● Miller Lee● Junior student at NCTU CS● Interests: C/C++, PL, CA, OS, compiler,

parallel programming, optimization

Why C++ AMP? ● GPUs can be 10+X

faster than CPUs for parallel code

● CUDA and OpenCL are still too complex/verbose for programmers

GPU computing

require explicit transfer

What we need in GPU programming

1. put data parallel codes into a kernel for GPU to execute

2. pass the arguments to GPU○ We can not pass the arguments by stack

3. an index to indicate current thread4. move the data between GPU and CPU

memory

OpenCL as an example

Device code in OpenCL__kernel voidmatrixMul(__global float* C, __global float* A, __global float* B, int wA, int wB){ int tx = get_global_id(0); int ty = get_global_id(1); float value = 0; for (int k = 0; k < wA; ++k) { float elementA = A[ty * wA + k]; float elementB = B[k * wB + tx]; value += elementA * elementB; } C[ty * wA + tx] = value;}

Host code in OpenCL 1.2

1. allocate and initialize memory on host side2. Initialize OpenCL3. allocate device memory and move the data4. Load and build device code5. Launch kernel

a. append arguments6. move the data back from device

intmain(int argc, char** argv){ // set seed for rand() srand(2006); // 1. allocate host memory for matrices A and B unsigned int size_A = WA * HA; unsigned int mem_size_A = sizeof(float) * size_A; float* h_A = (float*) malloc(mem_size_A); unsigned int size_B = WB * HB; unsigned int mem_size_B = sizeof(float) * size_B; float* h_B = (float*) malloc(mem_size_B); // 2. initialize host memory randomInit(h_A, size_A); randomInit(h_B, size_B); // 4. allocate host memory for the result C unsigned int size_C = WC * HC; unsigned int mem_size_C = sizeof(float) * size_C; float* h_C = (float*) malloc(mem_size_C); // 5. Initialize OpenCL // OpenCL specific variables cl_context clGPUContext; cl_command_queue clCommandQue; cl_program clProgram; size_t dataBytes; size_t kernelLength; cl_int errcode; // OpenCL device memory for matrices cl_mem d_A; cl_mem d_B; cl_mem d_C; /*****************************************/ /* Initialize OpenCL */ /*****************************************/ clGPUContext = clCreateContextFromType(0, CL_DEVICE_TYPE_GPU, NULL, NULL, &errcode); shrCheckError(errcode, CL_SUCCESS);

// get the list of GPU devices associated // with context errcode = clGetContextInfo(clGPUContext, CL_CONTEXT_DEVICES, 0, NULL, &dataBytes); cl_device_id *clDevices = (cl_device_id *) malloc(dataBytes); errcode |= clGetContextInfo(clGPUContext, CL_CONTEXT_DEVICES, dataBytes, clDevices, NULL); shrCheckError(errcode, CL_SUCCESS);

//Create a command-queue clCommandQue = clCreateCommandQueue(clGPUContext, clDevices[0], 0, &errcode); shrCheckError(errcode, CL_SUCCESS); // Setup device memory d_C = clCreateBuffer(clGPUContext, CL_MEM_READ_WRITE, mem_size_A, NULL, &errcode); d_A = clCreateBuffer(clGPUContext, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, mem_size_A, h_A, &errcode); d_B = clCreateBuffer(clGPUContext, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, mem_size_B, h_B, &errcode); // 6. Load and build OpenCL kernel char *clMatrixMul = oclLoadProgSource("kernel.cl", "// My comment\n", &kernelLength); shrCheckError(clMatrixMul != NULL, shrTRUE); clProgram = clCreateProgramWithSource(clGPUContext, 1, (const char **)&clMatrixMul, &kernelLength, &errcode); shrCheckError(errcode, CL_SUCCESS); errcode = clBuildProgram(clProgram, 0, NULL, NULL, NULL, NULL); shrCheckError(errcode, CL_SUCCESS); clKernel = clCreateKernel(clProgram, "matrixMul", &errcode); shrCheckError(errcode, CL_SUCCESS);

// 7. Launch OpenCL kernel size_t localWorkSize[2], globalWorkSize[2]; int wA = WA; int wC = WC; errcode = clSetKernelArg(clKernel, 0, sizeof(cl_mem), (void *)&d_C); errcode |= clSetKernelArg(clKernel, 1, sizeof(cl_mem), (void *)&d_A); errcode |= clSetKernelArg(clKernel, 2, sizeof(cl_mem), (void *)&d_B); errcode |= clSetKernelArg(clKernel, 3, sizeof(int), (void *)&wA); errcode |= clSetKernelArg(clKernel, 4, sizeof(int), (void *)&wC); shrCheckError(errcode, CL_SUCCESS); localWorkSize[0] = 16; localWorkSize[1] = 16; globalWorkSize[0] = 1024; globalWorkSize[1] = 1024; errcode = clEnqueueNDRangeKernel(clCommandQue, clKernel, 2, NULL, globalWorkSize, localWorkSize, 0, NULL, NULL); shrCheckError(errcode, CL_SUCCESS); // 8. Retrieve result from device errcode = clEnqueueReadBuffer(clCommandQue, d_C, CL_TRUE, 0, mem_size_C, h_C, 0, NULL, NULL); shrCheckError(errcode, CL_SUCCESS); // 10. clean up memory free(h_A); free(h_B); free(h_C); clReleaseMemObject(d_A); clReleaseMemObject(d_C); clReleaseMemObject(d_B); free(clDevices); free(clMatrixMul); clReleaseContext(clGPUContext); clReleaseKernel(clKernel); clReleaseProgram(clProgram); clReleaseCommandQueue(clCommandQue);

}

Nearly 200 lines of code

What is C++ AMP● C++ Accelerated Massive Parallelism

○ Designed for data level parallelism○ Extension of C++11 proposed by M$○ An open specification with multiple implementations

aiming at standardization■ MS Visual Studio 2013■ MCW CLAMP

● GPU data modeled as C++14-like containers for multidimensional arrays

● GPU kernels modeled as C++11 lambda

ComparisonsC++AMP Thrust Bolt OpenACC SYCL

Introsimple, elegant, performance(?), proposed by M$

library proposed by CUDA

library proposed by AMP

Annotation and

pragmas proposed

by SGI

wrapper for OpenCL proposed

by Codeplay

Matrix Multiplication in C++AMPvoid MultiplyWithAMP(int* aMatrix, int* bMatrix, int *productMatrix, int ha, int hb, int hc) { array_view<int, 2> a(ha, hb, aMatrix); array_view<int, 2> b(hb, hc, bMatrix); array_view<int, 2> product(ha, hc, productMatrix); parallel_for_each( product.extent, [=](index<2> idx) restrict(amp) { int row = idx[0]; int col = idx[1]; for (int inner = 0; inner < 2; inner++) { product[idx] += a(row, inner) * b(inner, col); } } ); product.synchronize();}

clGPUContext = clCreateContextFromType(0, CL_DEVICE_TYPE_GPU, NULL, NULL, &errcode); shrCheckError(errcode, CL_SUCCESS); // get the list of GPU devices associated // with context errcode = clGetContextInfo(clGPUContext, CL_CONTEXT_DEVICES, 0, NULL, &dataBytes); cl_device_id *clDevices = (cl_device_id *) malloc(dataBytes); errcode |= clGetContextInfo(clGPUContext, CL_CONTEXT_DEVICES, dataBytes, clDevices, NULL); shrCheckError(errcode, CL_SUCCESS); //Create a command-queue clCommandQue = clCreateCommandQueue(clGPUContext, clDevices[0], 0, &errcode); shrCheckError(errcode, CL_SUCCESS);

__kernel voidmatrixMul(__global float* C, __global float* A, __global float* B, int wA, int wB){ int tx = get_global_id(0); int ty = get_global_id(1); float value = 0; for (int k = 0; k < wA; ++k) { float elementA = A[ty * wA + k]; float elementB = B[k * wB + tx]; value += elementA * elementB; } C[ty * wA + tx] = value;}

Only 20 lines of codebut performance?

C++AMP programming modelvoid MultiplyWithAMP(int* aMatrix, int* bMatrix, int *productMatrix) { array_view<int, 2> a(3, 2, aMatrix); array_view<int, 2> b(2, 3, bMatrix); array_view<int, 2> product(3, 3, productMatrix); parallel_for_each( product.extent, [=](index<2> idx) restrict(amp) { int row = idx[0]; int col = idx[1]; for (int inner = 0; inner < 2; inner++) { product[idx] += a(row, inner) * b(inner, col); } } ); product.synchronize();}

GPU data modeled as data container

C++AMP programming modelvoid MultiplyWithAMP(int* aMatrix, int* bMatrix, int *productMatrix) { array_view<int, 2> a(3, 2, aMatrix); array_view<int, 2> b(2, 3, bMatrix); array_view<int, 2> product(3, 3, productMatrix); parallel_for_each( product.extent, [=](index<2> idx) restrict(amp) { int row = idx[0]; int col = idx[1]; for (int inner = 0; inner < 2; inner++) { product[idx] += a(row, inner) * b(inner, col); } } ); product.synchronize();}

Execution interface; marking an implicitly parallel region for GPU execution

C++AMP programming modelvoid MultiplyWithAMP(int* aMatrix, int* bMatrix, int *productMatrix) { array_view<int, 2> a(3, 2, aMatrix); array_view<int, 2> b(2, 3, bMatrix); array_view<int, 2> product(3, 3, productMatrix); parallel_for_each( product.extent, [=](index<2> idx) restrict(amp) { int row = idx[0]; int col = idx[1]; for (int inner = 0; inner < 2; inner++) { product[idx] += a(row, inner) * b(inner, col); } } ); product.synchronize();}

Kernels modeled as lambdas; arguments are implicitly modeled as captured variables

MCW C++AMP (CLAMP)

● Clang/LLVM-based○ translate C++AMP code to OpenCL C code and

generate OpenCL SPIR file○ With some template library

● Runtime support: gmac/OpenCL/HSA Okra● An Open Source project

○ The only two C++ AMP implementations recognized by HSA foundation (the other is MSVC)

○ Microsoft and HSA foundation supported

MCW C++ AMP Compiler

● Device Path○ generate OpenCL C code by

CBackend○ emit kernel function

● Host Path○ preparation to launch the

code

C++ AMP source code

Clang/LLVM 3.3

Device Code Host Code

Execution processC++ AMP

source code

Clang/LLVM 3.3

Device Code

C++ AMP source code

Clang/LLVM 3.3

Host Code

gmac

OpenCL

Our work

gmac

● unified virtual address space in software

● Can have high overhead sometimes

● In HSA (AMD Kaveri), GMAC is no longer needed

Compiling C++AMP to OpenCL

● C++AMP → LLVM IR → subset of C● arguments passing (lambda capture vs

function calls)● explicit V.S. implicit memory transfer ● Heavy works were done by compiler and

runtime

lambda capturestruct add { int a; add(int a) : a(a) {} int operator()(int x) const { return a + x; }};int main(void){ int a = 3; auto fn = [=] (int x) { return a + x; }; int c = fn(3); return 0;}

Those arguments should be put on the argument lists of OpenCL kernel.

What we need to do?

● Kernel function○ emit the kernel function with required arguments

● In Host side○ a function that recursively traverses the object and

append the arguments to OpenCL stack.● In Device side

○ reconstructor it on the device code for future use.

Examplestruct A { int a; };struct B : A { int b; };struct C { B b; int c; };struct C c;c.c = 100;auto fn = [=] () { int qq = c.c; };

Kernel code

__kernel void(int a, int b, int c){ C c(a, b, c); ...}

Deserialization constructorstruct C{ B b; int c; C (int a, int b, int c) : c(c), b(a, b) {}};

Serialization constructorstruct C{ B b; int c; void __cxxamp_serialize(Concurrency::Serialize s) { b.__cxxamp_serialize(s); s.Append(sizeof(int), &c); }};

Translationparallel_for_each(product.extent, [=](index<2> idx) restrict(amp) { int row = idx[0]; int col = idx[1]; for (int inner = 0; inner < 2; inner++) { product[idx] += a(row, inner) * b(inner, col); } });

__kernel voidmatrixMul(__global float* C, __global float* A, __global float* B, int wA, int wB){ int tx = get_global_id(0); int ty = get_global_id(1); float value = 0; for (int k = 0; k < wA; ++k) { float elementA = A[ty * wA + k]; float elementB = B[k * wB + tx]; value += elementA * elementB; } C[ty * wA + tx] = value;}

● Append the arguments● Set the index● emit kernel function● implicit memory management

Future work

● Future work for us○ restrict(auto)○ HSA related work

Future works for you

● Try this out!!● Many of us get spoiled and don’t want to go

back to write OpenCL directly anymore :-)● related links

○ Driver○ Clang○ sandbox